Remove the dead grok composer backend from swarm

.claude-plugin/marketplace.json

@@ -29,8 +29,8 @@
     {
       "name": "swarm",
       "source": "./plugins/swarm",
-      "description": "Local mixture-of-agents code review for Claude Code. Fans a diff across Claude lenses plus the codex and grok CLIs (grok-4.5 + composer models), merges by mechanism with cross-family consensus, verifies solo findings, and presents one ranked report. Optional --fix / --loop applies the findings you agreed with; --pr reviews a GitHub PR diff and posts the result. Skills: /swarm:review, /swarm:agents.",
-      "version": "0.4.2"
+      "description": "Local mixture-of-agents code review for Claude Code. Fans a diff across Claude lenses plus the codex and grok CLIs (grok-4.5), merges by mechanism with cross-family consensus, verifies solo findings, and presents one ranked report. Optional --fix / --loop applies the findings you agreed with; --pr reviews a GitHub PR diff and posts the result. Skills: /swarm:review, /swarm:agents.",
+      "version": "0.4.3"
     }
   ]
 }

.claude/knowledge/_index.md

@@ -14,7 +14,7 @@
 - `features/herdr-close-automation.md` — `/close` in herdr: cwd-tab teardown, plugin SessionEnd hook, the one TUI-exit primitive, detached self-exit onto idle
 - `features/herdr-tab-glyphs.md` — Task-state glyphs (`○ ● ◇ ◆ ✓`) on herdr tab names: `states` mode inside the self-contained renderer (no sibling), sync-vs-`--cached` PR refresh per caller, exact-cwd rename rules, soft pr-flow shim
 - `features/task-archiving-on-close.md` — `/close` archives (not deletes) the task file; adaptive commit + ff-push to main
-- `features/swarm-backend-adapter.md` — Verified codex/grok CLI facts (schema-enforced JSON, effort mapping, stdin hang, composer defensive-parse voice) behind `swarm`'s adapter script
+- `features/swarm-backend-adapter.md` — Verified codex/grok CLI facts (schema-enforced JSON, effort mapping, stdin hang, model-aware grok readiness) behind `swarm`'s adapter script
 - `features/swarm-review-pipeline.md` — `/swarm:review` P2 pipeline (skill↔Workflow wiring, family-consensus, minimal security) + P5 `--fix`/`--loop` (in-session orchestration between workflow runs; deterministic termination + close-out box in `loop-closeout.py`) + `--pr` (review a PR diff, post via `gh` under the user's identity so pr-flow's `claude`-author poll ignores it; publish path is deterministic `pr-post.py`: per-cell sanitizer + stale-head gate + post)
 
 ## Deployment

.claude/knowledge/features/swarm-backend-adapter.md

@@ -1,9 +1,9 @@
 ---
 title: "Swarm Backend Adapter Layer"
 createdAt: 2026-07-03
-updatedAt: 2026-07-15
+updatedAt: 2026-07-17
 createdFrom: "PR #21"
-updatedFrom: "session: 2026-07-15"
+updatedFrom: "PR #37"
 pluginVersion: 1.8.2
 prime: false
 reindexedAt: 2026-07-12
@@ -32,11 +32,13 @@ debugging round.
   `--output-last-message <file>` (stdout carries the agent transcript,
   stderr the progress log); grok prints a response **envelope** on stdout —
   the validated object is its `.structuredOutput` field.
-- **The adapter pins `-m grok-4.5`** — the schema-capable model. grok 0.2.101
-  renamed it from `grok-build` (same upstream pin-rename class as codex's
-  `gpt-5.6-terra`; verified drop-in: identical envelope/`structuredOutput`
-  shape, `--single` unchanged). `grok-composer-2.5-fast` still rejects
-  `--effort` ("does not support parameter reasoningEffort").
+- **The adapter pins `-m grok-4.5`** — the schema-capable model, and since
+  swarm 0.4.3 the *only* grok model it supports. grok 0.2.101 renamed it from
+  `grok-build` (same upstream pin-rename class as codex's `gpt-5.6-terra`;
+  verified drop-in: identical envelope/`structuredOutput` shape, `--single`
+  unchanged). Any other `--model` is preflight-rejected with a usage error —
+  only grok-4.5 enforces `--json-schema`, and an unlisted model fails late with
+  `structuredOutput: null` after burning a full review.
 - **Effort ladders**: grok is `low|medium|high` since 0.2.101 (the `max` tier
   is gone) → the adapter maps `xhigh`/`max`→`high`; codex has no `max` tier →
   map `max`→`xhigh` (`-c model_reasoning_effort=…`). Both mappings degrade a
@@ -47,23 +49,90 @@ debugging round.
   The pipeline runs codex at `high` normally; the `--max` profile overrides both
   model and effort (`gpt-5.6-sol` @ `xhigh`) — see
   [swarm-review-pipeline](swarm-review-pipeline.md).
-- **`grok-composer-2.5-fast` does not enforce `--json-schema`** — but it is
-  still usable as a second grok voice. Given a strict-JSON *prompt* (the adapter
-  appends the schema text and drops `--json-schema`/`--effort`), it emits **pure
-  `{"findings":[...]}` directly on stdout — no response envelope, no
-  `structuredOutput`** (verified P2, grok 0.2.82; simpler than grok-4.5's
-  envelope). The adapter routes `--model grok-composer-2.5-fast` to a separate
-  `run_grok_composer` path that parses the answer **defensively**: collect ALL
-  balanced `{...}` objects (whole string, fenced blocks, every `{` run), pick the
-  first **non-empty** `findings` object (a leading `{"note":…}`/`{"findings":[]}`
-  would otherwise mask the real one), and **validate every item against
-  `finding.schema.json`** (composer, unlike codex/grok-4.5, is not CLI-schema-
-  enforced — a malformed item must ERROR, not reach merge/verify). Both the
-  first-object bug and the missing per-item validation were caught by swarm
-  reviewing its own diff. composer is
-  ~2× faster than grok-4.5. It is same-family-correlated, so consumers must
-  count consensus by **model family**, not backend (composer + grok-4.5
-  agreeing is one grok vote — see [swarm-review-pipeline](swarm-review-pipeline.md)).
+- **Model-aware readiness beats an auth-only check** (swarm 0.4.3). grok drops
+  and renames models between releases — 0.2.101 removed
+  `grok-composer-2.5-fast`, which swarm had shipped as a second grok voice; the
+  auth-only probe kept reporting it Ready until it failed mid-review with
+  `Invalid params: "unknown model id"`. `ready`/`list` now also require
+  grok-4.5 in `grok models` (grok is the one backend with a usable model-list
+  command; codex has none, so its model is trusted). The gotchas, all live-
+  verified:
+  - **Parse the bullet list by SHAPE, not position — and match the id
+    SUBSTRING, not the field.** Lines read `  * grok-4.5 (default)`, but keying
+    on `$2` turns a reworded line (`  * default: grok-4.5`) into garbage tokens
+    that read as "the model is gone" and fail closed. Matching whole
+    whitespace-fields starting `grok-` fixes that but still drags glued-on
+    punctuation along (`grok-4.5,`, `` `grok-4.5` ``, ANSI codes), which breaks
+    the exact-match the same way. Extract with a pattern ending on
+    alphanumerics (`grok-[A-Za-z0-9]+([._-][A-Za-z0-9]+)*`); a line with no
+    id-shaped token then contributes nothing, landing in the trust-auth degrade.
+  - **Match without a pipe to `grep -q`**: an early-exiting `grep -q` can
+    SIGPIPE the writer, and under `set -o pipefail` a *hit* would then report
+    failure. Newline-fence the list and use a `case` substring match.
+  - **An empty model list must NOT fail closed.** Offline, a timeout, or a
+    future CLI renaming the subcommand would otherwise silently drop grok from
+    every fan-out. Empty/unparseable → trust auth and let `run_grok` surface the
+    explicit error; a non-empty list *without* grok-4.5 → an honest "not ready"
+    plus an update-the-CLI hint.
+  - **A probe added to a local path must not make it hang — and must not lie
+    when it can't run.** `ready`/`list` were purely local (stat the auth file)
+    before this; the probe puts a network call in every `/swarm:agents` and
+    review start. With no coreutils `timeout`/`gtimeout` to bound it (stock
+    macOS), the probe is **skipped** rather than run uncapped, degrading to
+    trust-auth in ~25ms — but it **warns on stderr**, because a silent skip
+    would make the documented model-aware guarantee false on that host: the
+    same "promise that doesn't hold at runtime" bug the composer removal exists
+    to fix.
+  - **Route every degrade through ONE audible exit.** This one spot was fixed
+    across FIVE consecutive swarm rounds, each catching the previous round's
+    miss: (1) the no-timeout branch ran uncapped; (2) it was capped but skipped
+    *silently*, while the docs were tightened to "never silently"; (3) `|| true`
+    still swallowed a *failed* probe, so the strengthened promise was false on
+    two of three routes; (4) with warnings finally on every route, `rc` was only
+    read **when the list came back empty** — so a probe killed mid-stream with
+    partial output skipped the degrade entirely and its truncated list read as
+    "model gone" (update an already-current CLI); (5) the probe was routed
+    through the review jail, which dragged `_init_sandbox`'s python3
+    profile-build into the local `ready`/`list` paths — where a missing python3
+    then misreported as "grok models failed". The invariants that survive:
+      - If N routes end in the same degrade, they need **one shared exit**
+        (`_probe_degraded`), not N hand-written warnings.
+      - **Check `rc` before the output, and discard partial output.** Only a
+        clean exit is an answer; anything else is a degrade.
+      - "The probe answered honestly" (model genuinely gone → `not ready` +
+        update-the-CLI hint) must stay distinguishable from "the probe never
+        answered" (→ trust auth + warning).
+    The meta-lesson (the one that actually ended the loop): **a fix that keeps
+    coming back is a shape problem, not a patch problem.** Rounds 1–4 patched a
+    probe that had a security jail bolted on; round 5 deleted the jail instead,
+    and the findings stopped. The composer removal itself — the PR's actual
+    subject — drew zero findings across all five rounds. When a *feature you
+    added to be safe* generates every round's bugs, cutting it beats hardening
+    it.
+  - **A readiness check is not a review — don't jail it.** The probe was first
+    built to go through `sandboxed()` (env-secret filter + sandbox-exec
+    read-deny jail), by analogy to the `run` calls. Wrong analogy: `sandboxed()`
+    exists because a *review* feeds grok the untrusted diff; a readiness check
+    passes **no** untrusted input, exactly like the sibling `codex login status`
+    a few lines away, which is also unjailed. Routing it through the jail bought
+    nothing and cost a hard python3 dependency on the formerly-local `ready`/
+    `list` paths (plus a shared-warning bug and a cross-backend memo bug — all
+    three vanished when the jail came back out). The probe runs grok directly.
+  - **Bound it with `timeout -k`, its own knob.** `SWARM_TIMEOUT` caps a
+    *review* (600s, `0` disables) — useless for a probe that `list` blocks on;
+    `SWARM_PROBE_TIMEOUT` (10s) is separate and a malformed/`0` value falls back
+    to 10, never uncapped. Plain `timeout` only SIGTERMs, so a grok that ignores
+    SIGTERM (or forks a stdout-inheriting child) keeps the `$(...)` substitution
+    blocking past the deadline — the "must never hang" hole. `-k <grace>` sends
+    SIGKILL after the grace period; treat both rc 124 (SIGTERM) and 137
+    (SIGKILL) as "timed out".
+  - **Memoize by call convention, not by wishing.** `list="$(grok_model_list)"`
+    runs the function in a *subshell*, so its cache-global assignments vanish
+    and every caller silently re-pays the network call. The cache only works if
+    the fetch is called directly and callers read the global
+    (`grok_model_fetch; local list="$_grok_models"`).
+  This mirrors work-system's `agent-registry.sh`, which learned the same lesson
+  at task-launch time.
 - **Headless tool execution**: both CLIs run read-only commands (e.g.
   `git diff`) without extra approval flags — codex inside `-s read-only`
   sandbox, grok headless `-p` auto-approves read-only tools. So lens prompts

.claude/knowledge/features/swarm-review-pipeline.md

@@ -1,9 +1,9 @@
 ---
 title: "Swarm Review Pipeline (/swarm:review)"
 createdAt: 2026-07-08
-updatedAt: 2026-07-15
+updatedAt: 2026-07-17
 createdFrom: "PR #24"
-updatedFrom: "session: 2026-07-15"
+updatedFrom: "PR #37"
 pluginVersion: 1.8.2
 prime: false
 reindexedAt: 2026-07-12
@@ -13,9 +13,11 @@ reindexedAt: 2026-07-12
 
 P2 turns the blueprint into a working review: a **Workflow-tool script**
 (`plugins/swarm/workflows/swarm-review.js`) launched by the `/swarm:review`
-skill. Shape: `scope+gate → fan-out (4 voices) → merge (file,mechanism) →
-verify solos → output-gated synthesis`. Four voices: Claude lenses ∥ codex ∥
-grok-4.5 ∥ composer (see [swarm-backend-adapter](swarm-backend-adapter.md)).
+skill. Shape: `scope+gate → fan-out (3 voices) → merge (file,mechanism) →
+verify solos → output-gated synthesis`. Three voices: Claude lenses ∥ codex ∥
+grok-4.5 (see [swarm-backend-adapter](swarm-backend-adapter.md)). A fourth,
+`grok-composer-2.5-fast`, was removed in swarm 0.4.3 — the grok CLI dropped the
+model.
 
 ## The skill ↔ workflow wiring (the non-obvious parts)
 
@@ -38,10 +40,16 @@ grok-4.5 ∥ composer (see [swarm-backend-adapter](swarm-backend-adapter.md)).
 
 ## Design decisions
 
-- **Consensus counts model *families*, not backends.** A cross-family cluster
+- **Consensus counts model *families*, not voices.** A cross-family cluster
   (≥2 of claude / openai / grok) is CONFIRMED without extra verify; everything
-  else is solo and goes through the adversarial 3-state verifier. composer +
-  grok-4.5 agreeing is one grok vote — they cannot alone mint consensus.
+  else is solo and goes through the adversarial 3-state verifier. The `FAMILY`
+  map (`workflows/swarm-review.js`) survives the composer removal deliberately:
+  with composer gone the backend→family mapping is 1:1, so *for consensus
+  counting* it is currently a no-op — but the fan-out is many-voices-per-family
+  (one Claude finder per gated lens), so the rule "same vendor agreeing with
+  itself is one vote, not a cross-check" is still the load-bearing invariant.
+  Collapsing consensus onto backends/voices would silently re-introduce
+  correlated self-agreement the day a second same-vendor voice returns.
 - **Security is intentionally minimal** (user directive: no cannons-at-sparrows).
   The P1 adapter floor stays (sandbox, tool-less grok, secret scrub, env filter,
   caps); P2 adds only three cheap things — **fencing** the diff as data
@@ -60,8 +68,8 @@ grok-4.5 ∥ composer (see [swarm-backend-adapter](swarm-backend-adapter.md)).
   the delimiter); the workflow only collision-checks it against the returned
   findings and extends it deterministically (`nonce-1`, `-2`…) on collision.
 
-- **`args.claude: false`** runs an **external-only control** (codex + grok-4.5
-  + composer, no Claude finder lenses, no gate; merge/verify still in-session).
+- **`args.claude: false`** runs an **external-only control** (codex + grok-4.5,
+  no Claude finder lenses, no gate; merge/verify still in-session).
   Proven useful: a control run found real bugs the with-Claude run missed (an
   `aws_secret_access_key` scrub-list drift, `git diff` omitting untracked files)
   — the "different models catch different defects" premise, live.
@@ -95,9 +103,9 @@ the diff out of the script, above). Claude applies edits between rounds.
   the report table contract this entry defines above (P2 reserved them).
 - **`--max` profile** (`INPUT.max` in the workflow): lifts every voice to its
   ceiling — codex `gpt-5.6-sol`@`xhigh` (codex has NO `max` tier, xhigh is its
-  top), Claude finder lenses + verifier `xhigh`; gate/merge, grok (`high` is
+  top), Claude finder lenses + verifier `xhigh`; gate/merge and grok (`high` is
   its ceiling since grok 0.2.101 dropped `max` — it runs there on both
-  profiles) and composer (no effort control) unchanged. Orthogonal to `--fix`/`--loop`,
+  profiles) unchanged. Orthogonal to `--fix`/`--loop`,
   composes with both. The profile's live settings are verified end-to-end
   (`gpt-5.6-sol`@`xhigh` at wiring time; grok re-verified at `--effort high` on
   0.2.101) — the "no silent fail on a non-existent model/effort" rule.
@@ -170,7 +178,9 @@ call, more thinking) over N calls. Not a default.
 
 Real background runs on this branch: a Claude-only smoke run proved the wiring
 (6 agents, correct return shape); the review **found a real bug in its own
-composer parser** (first-object-vs-findings-object), which was then fixed. Only
+grok-composer parser** (first-object-vs-findings-object), which was then fixed
+— that parser has since been removed with the composer backend (0.4.3), but it
+stands as the proof that swarm can catch a genuine bug in its own diff. Only
 `REFUTED` solos are dropped; consensus/solo/refuted counts + per-lens raw→
 surviving ship in the `balance` block the skill renders. Iterating a
 `/swarm:review` loop over the branch caught several fix-induced regressions

CHANGELOG.md

@@ -185,6 +185,10 @@ entries are grouped per plugin, newest first.
 
 ## swarm
 
+### 0.4.3 — 2026-07-17
+- Remove the `grok-composer-2.5-fast` backend: grok CLI 0.2.101 dropped the model, so the composer voice (adapter path, defensive parser, workflow voice, docs) failed at runtime. `grok-4.5` is now the only grok model, and the ensemble is three voices (Claude lenses + codex + grok-4.5).
+- Make grok readiness model-aware: `ready`/`list` now require `grok-4.5` to appear in `grok models`, not just auth — so a dropped or renamed model reads as "not ready" with an actionable hint instead of failing mid-review. The probe runs unjailed like the sibling `codex login status` check (a readiness check passes no untrusted diff, so it needs no sandbox and stays free of the review path's python3 profile-build), bounded by `SWARM_PROBE_TIMEOUT` (10s, `timeout -k` so a SIGTERM-ignoring CLI can't hang it; separate from the review-length `SWARM_TIMEOUT`). It falls back to the auth check — with a warning on stderr, never silently — whenever it can't produce a clean answer (no coreutils `timeout`, non-zero exit, timeout, or an empty/unparseable list), rather than dropping grok from the fan-out.
+
 ### 0.4.2 — 2026-07-15
 - Fix grok CLI 0.2.101 compat: pin `grok-4.5` (upstream renamed `grok-build`), cap grok effort at `high` (the `max` tier is gone; the adapter maps `xhigh`/`max` → `high` so stale callers degrade instead of erroring).
 

CLAUDE.md

@@ -21,7 +21,7 @@ This is a **Claude Code plugin marketplace** (monorepo) containing plugins that
 - **knowledge-system** (v1.9.x) — Knowledge management with three layers: Rules, Knowledge, Memory. Skills: `/init`, `/query`, `/curate`, `/reindex`, `/backfill-knowledge`, `/migrate`, `/statusline`
 - **work-system** (v1.8.x) — Task and worktree workflow. Skills: `/define`, `/kickoff`, `/adopt`, `/continue`, `/status`, `/close`, `/list`, `/statusline`
 - **pr-flow** (v1.3.x) — PR review feedback loop. Skills: `/open`, `/cycle`, `/check`, `/fix`, `/rebase`, `/merge`
-- **swarm** (v0.4.x) — Local mixture-of-agents code review (external `codex`/`grok` CLIs — grok-4.5 + composer — plus Claude lenses). P2: `/swarm:review` pipeline (scope→fan-out→merge→verify); P5: `--fix`/`--loop` apply the findings you agreed with. Skills: `/swarm:review`, `/swarm:agents`
+- **swarm** (v0.4.x) — Local mixture-of-agents code review (external `codex`/`grok` CLIs — grok-4.5 — plus Claude lenses). P2: `/swarm:review` pipeline (scope→fan-out→merge→verify); P5: `--fix`/`--loop` apply the findings you agreed with. Skills: `/swarm:review`, `/swarm:agents`
 
 ## Plugin Anatomy